Google unveils multimodal Gemma 4

Google DeepMind has unveiled the new Gemma 4 model series under the Apache 2.0 license. The lineup consists of models in 2B, 4B, 31B sizes and a 26B-A4B Mixture-of-Experts (MoE) model.

Key Features and Technology:

• Multimodal capabilities: All models natively process video and images, and in particular the E2B and E4B models support native audio input for speech recognition.
• Maximizing parameter efficiency: The smaller models (E2B, E4B) apply Per-Layer Embeddings (PLE) technology. This assigns unique embeddings to each decoder layer, designed to deliver strong performance in on-device environments while keeping the actual parameter count low.
• High intelligence density: Google focused on maximizing performance relative to parameter count, delivering strong reasoning capabilities even in the smaller models.